Compute Shard is an experimental distributed inference system that splits a 120B parameter model across multiple scattered consumer GPUs. It features a shared terminal where users can collectively watch the swarm stream output in real-time.